Hi, Junzhi
Thank you for your meticulous study of JVM default parameters. It's really cool.
Given that IoTDB is applicable to many scenarios, including edge and cloud
sides, adjusting default parameters is always challenging to benefit all
scenarios. I support the current choice of parameters, commenting out some
parameters whose benefits and harms are uncertain, but explaining through
comments the scenarios in which they are beneficial. Commit parameters that are
definitively beneficial and harmless.
Additionally, further introduction is needed on which logs the SafepointTimeout
parameter should be printed to, which can be added to the comments~
Thanks
------------------------
Xinyu Tan
On 2024/02/21 04:29:06 Junzhi Peng wrote:
> Hi all,
>
> We have recently investigated some GC options. Based on the product
> scenarios of IoTDB, we have selected some general JVM GC options. For
> options that are completely beneficial and harmless, we plan to
> directly add them to the default startup options of IoTDB; For options
> that are only beneficial in some specific scenarios, we plan to add
> them to default startup options through comments and explain their
> usage. Please review~
>
> 1. Directly added options:
>
> -Xss512k： This option is intended to optimize the memory occupied by
> threads. The default value is 1m. Based on our research and actual
> usage scenarios of iotdb, we suggest that the thread stack size of
> 512k is sufficient. Many online business systems also set up in this
> way.
>
> -XX:+UnlockDiagnosticVMOptions： Used to unlock some diagnostic JVM
> options, such as -XX:GuaranteedSafepointInterval. This option itself
> has no direct impact on performance.
>
> -XX:+UseCountedLoopSafepoints： Prevent the delay in entering the safe
> point caused by large bounded loops from causing the GC STW time to be
> too long.
>
> -XX:+SafepointTimeout -XX:SafepointTimeoutDelay=1000： These two
> options will monitor the waiting time of the jvm threads in safepoint.
> If it exceeds 1000ms, the state of the corresponding threads will be
> printed to the log to help us troubleshoot the problem of too long
> safepoint stw time.
>
> -XX:GuaranteedSafepointInterval=0： Disable scheduled safepoint tasks.
> In high concurrency scenarios of IoTDB, it is not necessary to enter
> the safepoint regularly, but may cause additional overhead.
>
> -XX:-UseBiasedLocking： Disable biased lock. In high concurrency
> scenarios of IoTDB, biased locks are basically useless and have
> negative effects with safepoints. In fact, biased locking has been
> disabled by default since JDK15.
>
> -XX:+UseAdaptiveSizePolicy： Turn on GC adaptive tuning（Ergonomics）
>
> 2. Added through comments:
>
> -XX:+ParallelRefProcEnabled： Turn on parallel processing of Reference
> objects in GC process. This option may help if particularly long
> reference processing times are observed in the GC logs. Please note:
> because this option consumes thread resources, it may have an impact
> on application's throughput.
>
> CPU_PROCESSOR_NUM=$(nproc) -XX:ParallelGCThreads=${CPU_PROCESSOR_NUM}：
> Set the processing thread of parallel gc to the number of machine CPU
> cores. Please note: When the GC time is too long, if there are
> remaining CPU resources, you can try to increase this option. Although
> this will occupy application thread resources, but because the GC
> becomes faster, it will have minor impact on the application. If CPU
> resources have reached a bottleneck, increasing this option may make
> application slower even if it makes the GC faster.
>
> -XX:+AlwaysPreTouch： Force the operating system to allocate all memory
> in advance when the JVM starts, so that all heap areas are allocated
> physical memory. It can reduce the overhead caused by memory
> allocation, page fault interrupts, etc. during JVM operation. In some
> large memory scenarios, we have observed about 10% improvement in
> throughput with this option. Please note: The use of this option will
> cause the JVM to start slower. At the same time, the physical memory
> usage of iotdb will reach the maximum heap memory immediately after
> JVM startup, which may reduce memory utilization and trigger OOM
> killer when memory is tight.
>
>
> Best regards,
> —————————————
> Junzhi Peng
>
